Changelog

All notable changes to Gia (Gentari AI Assistant) will be documented in this file.

v1.2 - Grounded Answers

Grounding Verification - Gia now drafts an answer, verifies it against retrieved document excerpts, and retries if the draft contains unsupported claims.
Safe Fallbacks - If Gia still cannot verify an answer from the selected document(s), it says so instead of guessing.

Stricter Document Prompting - The assistant prompt now treats retrieved excerpts as the only source of truth and explicitly avoids filling gaps with general knowledge.

Localhost Startup - Local runs now bind to 127.0.0.1 and auto-select the next free port when another process is already using localhost:5200.

All Documents Default - Chat now defaults to searching across all ready documents.
Cross-Document Retrieval - Gia ranks matches across every ready document and shows source chips for the documents used.
Bulk Uploads - Admin can upload multiple files in one action.
Expanded File Support - Added parsing for Office files, spreadsheets, presentations, OpenDocument formats, EPUB, EML, CSV/TSV, JSON/JSONL, XML/HTML/SVG, RTF, YAML, code/config files, and existing PDF/image OCR flows.

Safer Ingestion Queue - Document ingestion now runs through a single background queue to avoid native-library contention during large uploads.
FAISS Stability - Vector stores now default to FlatIP indexes with FAISS/OpenMP threads capped to 1 on macOS.
Default Port - Local app defaults and deployment docs now use port 5200.

Bulk Upload Crash - Fixed a macOS segmentation fault caused by concurrent FAISS HNSW index creation.
Metadata Writes - Document metadata is now written atomically to reduce corruption risk after crashes.

AI-Powered Chat — Conversational RAG assistant powered by Gemma 3 (1B) via Ollama
Multi-Model Support — Upload and switch between multiple knowledge models
PDF Parsing — 3-tier parsing engine: Marker → PyMuPDF + Tesseract OCR → pypdf
Image OCR — Extract text from PNG, JPG, TIFF, BMP via Tesseract
FAISS Vector Store — Per-model semantic search with nomic-embed-text embeddings
Streaming Responses — Real-time token-by-token SSE streaming
Session Management — Persistent chat sessions with localStorage
Admin Dashboard — Upload, manage, and delete models with ingestion status tracking
Health Monitoring — Live system status panel (Server, Ollama, Models, LLM)
Modern UI — Clean glassmorphism design with sidebar navigation
Responsive Design — Mobile-friendly layout with collapsible sidebar